GPT-6 solved a Millennium problem—does your team need that?

OpenAI's GPT-6 Astra, trained on 100,000 Blackwell systems, achieved 95.9% on a CAD benchmark and reportedly solved the Navier-Stokes Millennium Prize problem using 10,000 AI agents coordinated by humans. The solution took 88 hours and required tens of millions in compute, yet the mathematical community must still verify the proof.

This raises a fundamental question: is the cost of solving rare problems with massive compute justified for enterprise use cases? A security startup used AI to discover a memory corruption vulnerability in WeChat's VoIP stack, creating a worm that spreads without user interaction. While the exploit was responsibly disclosed

and patched, the incident signals that AI makes sophisticated hacking more accessible, not more powerful. The real risk isn't new attack surfaces—it's that AI lowers barriers for adversaries to find vulnerabilities in software that can never be perfectly secure. IBM's US Open app demonstrates practical applied AI: real-time match

prediction using gradient-boosted models, conversational match chat with three competing agent architectures, and limb tracking with 21 points at 50fps across 254 matches. This shows enterprise AI working today at scale with known constraints. The contrast between speculative breakthroughs and production deployments highlights the gap engineering leaders must manage.